
M2P Fintech
Fintech is evolving every day. That's why you need our newsletter! Get the latest fintech news, views, insights, directly to your inbox every fortnight for FREE!

For most of the last decade, the core banking conversation centred on efficiency. Banks replaced ageing systems to cut operating costs and launch products faster. That work still matters, but it no longer decides on its own which institutions grow.
Artificial intelligence has added a second question. It is no longer enough for a core banking system to record transactions accurately. Banks now need a core that can feed models with clean, real-time data, act on model decisions safely, and expose products to the new channels where customers, and increasingly their AI assistants, make financial choices.
This blog looks at what AI means for core banking in practical terms. It covers where AI fits in the core banking stack, how banks around the world are using it, the benefits and the trade-offs, how regulators are responding, and who is accountable when an AI-driven decision goes wrong. It closes with a risk mitigation framework and the capabilities an AI-ready core banking platform needs.
AI has been part of banking for years, mostly in fraud scoring and credit models that sat at the edges of the core. Three shifts are now pulling it into the centre of the stack.
The McKinsey Global Institute estimates that generative AI alone could add $200 billion to $340 billion in annual value to global banking, equal to 9 to 15 percent of operating profits, mostly through productivity gains (McKinsey). Most of that value depends on access to accurate, well-structured data about customers, accounts, products and transactions. That data lives in the core.
Banking products are not losing relevance. Demand for deposits, loans and payments is steady. What is changing is where customers find them.
McKinsey estimates that embedded finance channels could initiate 20 to 25 percent of retail banking sales to individuals and SMEs by 2030 (McKinsey). Juniper Research forecasts global open banking API calls to rise from 137 billion in 2025 to 722 billion by 2029, and expects AI integration to accelerate that growth (Juniper Research).
Customers now begin many financial decisions at a checkout page, a marketplace, an accounting app or a search result. Increasingly, they begin with a conversation with an AI assistant.
In an Adobe survey, 38 percent of US consumers had already used generative AI for online shopping, and 73 percent of those users called it their main source of product research. Adobe also recorded a 4,700 percent year-on-year rise in generative AI referral traffic to US retail sites in July 2025, though that traffic was still small next to established channels (Adobe).
Financial services are likely to follow a similar path. An AI assistant asked to park idle cash will compare yields, lock-ins, fees and eligibility across providers. It can only shortlist products whose terms are available as structured, machine-readable data. A bank whose core cannot expose that data may simply not appear among the options.
Together, these shifts make AI a core banking question. Intelligence depends on the core's data, and distribution through AI channels depends on what the core can expose, price and control in real time.
AI does not replace the core banking system. The core remains the system of record: it holds balances, posts transactions, applies interest and fees, and enforces product rules. AI works around and on top of it, reading its data, recommending or taking decisions, and sending instructions back through controlled interfaces.
A useful way to plan AI adoption is to map it to the layers of the core banking stack.
|
Core banking layer |
What AI does here |
Typical examples |
|---|---|---|
|
Product and pricing engine |
Recommends product configurations, prices and limits using demand, risk and behaviour data |
Dynamic deposit rates, risk-based loan pricing, personalised bundles |
|
Customer and onboarding (CIF, KYC) |
Reads documents, matches identities and flags anomalies |
Document extraction, face match, duplicate-customer detection, video KYC assistance |
|
Accounts and transaction processing |
Monitors flows in real time and classifies transactions |
Fraud scoring, transaction categorisation, anomaly alerts |
|
Lending and credit |
Scores applicants, monitors portfolios and predicts stress |
Alternative-data underwriting, early-warning signals, collections prioritisation |
|
Risk, compliance and AML |
Screens, prioritises and explains alerts |
Sanctions screening, AML alert triage, suspicious-activity narratives, regulatory report drafting |
|
Treasury and liquidity |
Forecasts cash positions and deposit behaviour |
Liquidity forecasting, ALM scenario analysis |
|
Operations and back office |
Automates manual, repetitive tasks |
Reconciliation, exception handling, dispute resolution, end-of-day checks |
|
Engineering and change |
Speeds up configuration, testing and migration |
Code generation, automated test cases, data mapping for migrations |
|
Customer engagement |
Answers questions and guides customers |
Conversational assistants, relationship-manager copilots, next-best-action |
|
Distribution and agent interfaces |
Makes products discoverable and callable by external AI agents |
Structured product catalogues, APIs exposed as callable tools (for example, through MCP), consent and mandate controls |
Two patterns sit across all these layers.
AI as an advisor. The model recommends, and a human or a rules engine decides. This is common for credit decisions above a threshold, AML escalations and complaint handling.
AI as an actor. The model initiates an action within limits set by the bank, such as blocking a suspicious card, moving a payment to review or triggering a collection message. Agentic AI extends this further: an agent can run a multi-step workflow across systems, such as gathering documents, checking eligibility and opening an account.
The further a bank moves from advisor to actor, the more the core must enforce. Limits, permissions, audit trails and the ability to reverse an action need to be built into the core itself, not left to the AI layer.
Banks that report measurable AI value share one trait: they built clean data foundations and connected AI to core systems, rather than running isolated pilots. The examples below show how that plays out across regions.
|
Region |
Institution |
Use case |
Reported outcome |
|---|---|---|---|
|
United States |
In-house, model-agnostic generative AI platform (LLM Suite) and AI-based fraud detection |
200,000+ employees with access; fraud costs held flat despite a 12% annual rise in attacks |
|
|
United States |
Erica, an AI virtual financial assistant for retail, wealth and corporate clients |
3.2 billion+ client interactions since its 2018 launch |
|
|
United Kingdom |
Generative AI added to the Cora digital assistant, plus AI in wealth document processing and software engineering |
Generative AI-supported customer journeys grew from 4 to 21 in 2025 |
|
|
Singapore |
430+ AI use cases across customer engagement, operations and risk, run on 2,000+ models |
About S$1 billion in economic value from data analytics and AI/ML in 2025 |
|
|
India |
MuleHunter.AI, a shared machine learning model to detect mule accounts used in digital fraud |
Built on 19 mule-account behaviour patterns; reported as live in 26 banks and expanding |
|
|
UAE |
Agentic AI across trade, payments, client operations and compliance |
30+ agentic use cases in the production pipeline; 90%+ of structured data integrated on one platform |
|
|
UAE |
Generative AI models connected to internal systems, APIs and enterprise data, deployed on in-country infrastructure |
||
|
Kenya |
Alternative-data credit scoring using mobile top-ups, transaction histories and repayment patterns |
80%+ of loans disbursed digitally, often in under a minute |
Fraud and financial crime come first. Fraud detection is usually the fastest route to measurable return because losses avoided are easy to count. India's MuleHunter.AI also shows a different model: a regulator-backed tool that gives smaller banks access to capabilities they could not build alone.
Credit is where AI widens access. In Kenya, a 2025 Central Bank of Kenya survey found that 65 percent of banks using AI apply it to credit risk scoring (Dawan Africa). Mobile-money data has made it possible to lend to customers with no formal credit history.
Customer assistants scale once they connect to account data. Erica and Cora deliver value because they can read balances and transactions and act on them, not just answer FAQs.
The Middle East is moving straight to agentic AI. UAE banks are deploying agents that run multi-step processes, supported by unified data platforms and in-country model hosting that addresses data residency requirements.
Measurement is a discipline. DBS, for example, attributes economic value to individual use cases. That makes it easier to fund what works and stop what does not.
When AI is connected to a modern core, the gains show up in cost, risk, revenue and customer experience at the same time.
Faster product launches. AI can analyse demand and risk data to suggest product parameters, pricing bands and eligibility rules. Combined with a parameterised core, banks can move from idea to live product in weeks rather than quarters.
Lower operating costs. Reconciliation, exception handling, document checks and report drafting are repetitive and rule-heavy. Automating them frees operations teams for work that needs judgement.
Stronger fraud and AML controls. Machine learning models catch patterns that static rules miss, such as networks of mule accounts or unusual sequences of small transfers. Better models also reduce false positives, which cuts investigation workload and avoids blocking genuine customers.
Wider and fairer credit access. Alternative data, such as cash flow, GST filings, UPI history or mobile-money activity, lets lenders assess thin-file borrowers. Used responsibly, this brings first-time borrowers and small businesses into formal credit.
Better portfolio health. Early-warning models flag accounts likely to miss payments, so banks can restructure or engage before an account becomes non-performing.
Personalised engagement at scale. AI can recommend the next relevant product, nudge customers before a shortfall or suggest savings moves. This turns the core's transaction data into everyday value for the customer.
Faster, safer modernisation. Generative AI helps with code analysis, data mapping and test case generation during core migrations. This reduces the time and risk of moving off legacy systems.
Readiness for new distribution channels. A core that exposes products as structured data and callable APIs can be discovered and used by partner platforms and AI agents. This keeps the bank on the shelf as customer journeys move outside bank-owned channels.
The same features that make AI useful, such as learning from data and acting at speed, also create new failure modes. Banks need to plan for them from day one.
Bias and unfair outcomes. Models learn from historical data, and historical lending decisions may reflect past discrimination. A model can deny credit to a group without any explicit rule doing so, through proxies such as location or device type.
Opaque decisions. Complex models can be hard to explain. If a bank cannot tell a customer or an examiner why a loan was declined or an account frozen, it has a conduct and a compliance problem.
Hallucinations and errors in generative AI. Large language models can produce confident but wrong answers. In a customer assistant, a wrong statement about fees, rates or eligibility can mislead customers and create liability.
Model drift. Customer behaviour, fraud patterns and economic conditions change. A model trained on last year's data can quietly lose accuracy unless it is monitored and retrained.
Automation without affordability checks. Speed can outrun prudence. In Kenya, where instant mobile loans are common, central bank data showed default rates above 80 percent on loans below KES 1,000, and regulators have moved to require lenders to verify repayment capacity (TechCabal).
Data privacy and consent. AI needs large volumes of personal and transaction data. Using that data beyond the purpose for which it was collected, or sending it to external model providers without controls, can breach data protection laws.
New cyber attack surfaces. AI systems can be targeted through prompt injection, data poisoning and model theft. Attackers also use AI themselves, for deepfake voice and video in social engineering and for synthetic identities at onboarding.
Third-party and concentration risk. Many banks rely on a small number of cloud and model providers. An outage, a pricing change or a security incident at one provider can affect many institutions at once. India's central bank has publicly flagged this concentration risk (Inc42).
Legacy cores limit what AI can do. Batch-based cores with fragmented data cannot feed real-time models or act on their outputs instantly. Many AI pilots stall not because the model fails but because the core cannot support it.
Cost, skills and change management. Data engineering, model governance, monitoring and staff training all cost money. Workforce changes also need careful handling, including reskilling staff for oversight roles.
Agentic AI raises the stakes. When an AI agent can move money or open accounts, an error is no longer a bad recommendation. It is a completed transaction. Mandates, limits and reversibility become essential.
Regulators are not trying to stop banks from using AI. They are making it clear that a bank stays fully accountable for every decision its AI makes, whether it built the model or bought it. Approaches differ by country, but the direction is the same.
|
Jurisdiction |
Key instrument |
Status (as of November 2026) |
What it means for banks |
|---|---|---|---|
|
India |
RBI FREE-AI Committee Report; draft Model Risk Management guidance |
FREE-AI report published August 2025; MRM draft released June 2026 for consultation |
Board-approved AI and model risk policies, human oversight, kill-switch arrangements, explainability |
|
European Union |
In force; high-risk obligations for credit scoring now apply from 2 December 2027 |
Credit scoring is high-risk: risk management, documentation, human oversight, conformity assessment |
|
|
Singapore |
Issued 7 October 2026; phased effect from October 2027 |
AI inventories, lifecycle controls, third-party AI risk management; covers generative AI and AI agents |
|
|
United Kingdom |
No AI-specific rules; existing frameworks apply |
Consumer Duty for outcomes; senior managers personally accountable under SM&CR |
|
|
United States |
Replaced SR 11-7 on 17 April 2026 |
New model risk guidance excludes generative and agentic AI, but fair lending, consumer protection and third-party rules still apply |
The RBI's FREE-AI Committee set out seven guiding principles, or sutras: Trust, People First, Innovation over Restraint, Fairness, Accountability, Explainability and Resilience. These are backed by 26 recommendations across six pillars: infrastructure, policy, capacity, governance, protection and assurance (RBI; KPMG). The report also proposed an AI innovation sandbox and, in the medium term, a graded liability framework for AI (Mondaq).
In June 2026, the RBI followed with draft guidance on model risk management. It proposes a board-approved framework covering all models, including AI and models bought from vendors, with human oversight of automated decisions and arrangements to suspend or shut down a model if needed (ThePrint). Banks should track the final version of this guidance as it is issued.
The EU AI Act classifies AI used for creditworthiness assessment and credit scoring as high-risk. Under the Digital Omnibus amendments approved in 2026, these obligations now apply from 2 December 2027 instead of August 2026 (SecurePrivacy). Banks using such systems will need risk management, data governance, technical documentation, human oversight and conformity assessment. Rules on prohibited practices and general-purpose AI models already apply.
MAS issued its Guidelines on AI Risk Management in October 2026. They apply to all financial institutions and all forms of AI, including generative AI and AI agents. MAS expects institutions to maintain AI inventories and apply controls for data governance, cybersecurity, testing, human oversight, monitoring and change management (Fintech News Singapore). MAS treats AI as something that amplifies existing financial, operational, conduct and reputational risks, and places strong emphasis on third-party AI (HLC).
The FCA has said it will not create AI-specific rules. It relies on the Consumer Duty, the Senior Managers and Certification Regime (SM&CR) and its governance expectations (FCA). In practice, this means a named senior manager is accountable for AI outcomes in their area. The FCA also runs an AI Live Testing programme where firms test AI in real markets under supervision.
In April 2026, the Federal Reserve, OCC and FDIC replaced the 2011 model risk guidance (SR 11-7) with SR 26-2. The new guidance excludes generative and agentic AI from its scope, but fair lending, consumer protection, third-party risk, cybersecurity and recordkeeping obligations still apply to any AI a bank deploys (Anaconda). Under fair lending rules, lenders must still give specific reasons when they decline credit, regardless of how complex the model is.
Despite different legal styles, regulators converge on the same expectations:
Responsible AI is not a separate project from compliance. It is the set of choices that decide whether customers can trust a bank's automated decisions. Seven principles cover most of what banks need.
1. Fairness. Test models for disparate outcomes across gender, age, geography, caste, religion and other protected or sensitive characteristics, even when those attributes are not model inputs. Proxy variables such as PIN code or device type can reintroduce bias. Fairness testing should happen before launch and continuously after it.
2. Transparency. Customers should know when they are dealing with an AI system, especially in conversational channels. Material decisions, such as a loan rejection or an account restriction, should come with clear, specific reasons in plain language.
3. Explainability. Choose the simplest model that does the job. Where complex models are justified, use explanation techniques that show which factors drove a decision, and document them so that a reviewer can reproduce the reasoning.
4. Human agency and oversight. Keep humans in the loop for high-impact decisions, and on the loop for high-volume automated ones. Staff must have the authority, the information and the time to override a model, not just approve it.
5. Privacy and data minimisation. Use only the data needed for the purpose, with valid consent. Keep sensitive data within approved environments, and do not send customer data to external model providers without contractual and technical safeguards. In India, this means aligning with the Digital Personal Data Protection Act, 2023, alongside RBI data localisation requirements.
6. Inclusion and customer benefit. AI should widen access, not exclude. Alternative-data credit models should help thin-file borrowers in, and affordability checks should stop automation from pushing unsuitable credit. Customer-facing AI should support local languages and accessibility needs.
7. Accountability and redress. Every AI system needs a named owner. Customers must have a simple way to challenge an automated decision and reach a human who can review and reverse it.
A practical test helps: would the bank be comfortable explaining the decision, and how it was made, to the customer, the regulator and a journalist? If not, the design needs another look.
The short answer is that the bank is responsible. An AI model is not a legal person, and no regulator accepts "the algorithm decided" as a defence. What changes with AI is how responsibility is shared inside the bank and with its vendors, and how quickly a failure must be detected and fixed.
If an AI system wrongly declines a loan, freezes an account, misprices a product or gives a customer incorrect information, the customer's claim is against the bank. A well-known example outside banking is the 2024 Canadian tribunal ruling against Air Canada, which was held responsible for incorrect refund information given by its website chatbot. The same logic applies to a bank's virtual assistant: what the AI tells a customer is treated as what the bank told the customer.
Existing customer protection rules continue to apply. In India, for example, the RBI's framework on customer liability for unauthorised electronic transactions limits a customer's liability where the fault lies with the bank, regardless of whether a human or a model made the error.
Regulators are pushing accountability to named individuals:
Most banks use third-party models, platforms and data. Regulators are clear that outsourcing the technology does not outsource the responsibility. Banks can, however, allocate financial liability through contracts.
|
Failure scenario |
Accountable to the customer |
How the bank manages vendor exposure |
|---|---|---|
|
Biased credit model denies loans unfairly |
The bank |
Warranties on training data and testing; audit rights; obligation to support fairness reviews |
|
Chatbot gives wrong fee or eligibility information |
The bank |
Service levels on accuracy; content controls; indemnities for vendor-caused errors |
|
Fraud model misses an attack or blocks genuine customers |
The bank |
Performance thresholds; incident reporting timelines; shared monitoring dashboards |
|
AI agent executes an unauthorised or incorrect transaction |
The bank (and possibly the agent provider, depending on the mandate) |
Clear mandate and consent records; transaction limits; reversal procedures; liability terms with agent platforms |
|
Customer data leaks through an external model provider |
The bank, as data fiduciary |
Data processing terms; data residency; security certifications; breach notification and indemnity clauses |
When a customer's own AI agent instructs a bank to move money, responsibility may be split between the customer who set up the agent, the company that built it and the bank that executed the instruction. Laws have not settled this yet. Until they do, banks should rely on strong mandates: explicit customer consent, scoped permissions, transaction limits, step-up authentication for high-risk actions and a complete audit trail of what the agent asked for and what the bank did.
The safest AI programmes treat risk management as part of the build, not a review at the end. The framework below follows an AI system through its lifecycle.
|
Risk |
Primary controls |
|---|---|
|
Bias and unfair outcomes |
Pre-launch and ongoing fairness testing; proxy-variable review; diverse training data; human review of declines |
|
Opaque decisions |
Simpler models where possible; explanation tools; reason codes in decision logs; customer-facing explanations |
|
Generative AI errors |
Retrieval from approved bank content only; output filters; confidence thresholds; human handoff; regular accuracy audits |
|
Model drift |
Real-time performance monitoring; scheduled revalidation; automatic alerts and fallback |
|
Unsafe autonomous actions |
Scoped permissions; transaction limits; mandates and consent records; reversibility; full audit trails |
|
Data privacy breaches |
Data minimisation; encryption; in-country hosting; contractual controls on model providers |
|
Cyber attacks on AI |
Prompt-injection defences; access controls; adversarial testing; deepfake-aware authentication |
|
Vendor and concentration risk |
Multi-model architecture; exit plans; vendor audits; resilience testing |
Banks cannot apply AI to data they cannot reach, or act on decisions their core cannot execute. An AI-ready core is defined less by the AI it ships with and more by how well it feeds, executes and controls AI.
Real-time data and events. Models need current balances, transactions and customer activity, not yesterday's batch file. An event-driven core publishes every change as it happens, so fraud, credit and engagement models can respond in milliseconds.
API-first architecture. Every core capability, from account opening to loan disbursement, should be available through secure, documented APIs. This lets AI services read data and trigger actions without custom integrations for each use case.
Composability. Banks should be able to assemble and vary products from reusable components by channel, partner and segment, without building a separate stack for each. Composability also lets modernisation happen in stages rather than as one large migration.
Cloud-native scale. AI-driven and partner-led distribution creates sudden spikes in demand. A cloud-native core scales capacity up and down and supports new partner arrangements without a fresh infrastructure programme each time.
Agent-ready interfaces. As AI assistants start comparing and buying financial products, banks will need product terms published as structured data and banking capabilities exposed as callable tools, for example through the Model Context Protocol (MCP). These must sit behind mandate, consent and audit controls so the bank can act safely on an agent's instruction.
Built-in controls. Limits, maker-checker workflows, role-based permissions, partner-level access controls and complete audit trails must be enforced in the core itself. That way, no AI layer can bypass them.
Clean, unified data. A single customer view and consistent product and transaction data reduce the effort of training and validating models, and make AI decisions easier to explain.
Security and compliance by design. Data residency, encryption, consent management and real-time monitoring need to be part of the platform, because AI and distributed banking expand the surface area for fraud, third-party risk and regulatory failure.
AI delivers value in banking only when the core can feed it real-time data, act on its decisions and keep every action within the bank's controls. M2P's Turing Core Banking System is built for exactly that.
Turing CBS is a cloud-native, API-first core banking system with a real-time, event-driven ledger. It can replace a legacy core or run alongside it, so banks can modernise in stages rather than through a single big-bang migration (M2P).
Real-time data for every model. The event-driven ledger has no batch windows or end-of-day bottlenecks, so fraud, credit and engagement models work on live balances and transactions.
AI agents built into the core. Purpose-built agents run inside the system of record rather than as a separate tool on top:
Because these agents operate inside the core, they inherit its permissions, limits and audit controls.
Controls that AI cannot bypass. Built-in regulatory reporting, audit trails and maker-checker workflows are enforced in the core. This supports the board-level accountability, human oversight and traceability that regulators such as the RBI and MAS now expect.
Faster, configurable products. A universal product factory lets banks configure savings, current, term deposit and wallet products without a code release, so AI-informed product and pricing decisions reach the market quickly.
One data model across banking, lending and payments. Turing CBS sits within M2P's wider stack of 25+ products, connected through Connect, M2P's orchestration layer. Vue by M2P, the shared intelligence layer, adds document, decision, operational and risk intelligence. A signal captured once, such as a KYC document or a UPI fraud pattern, can inform decisions across products.
Built for complex institutions. Multi-entity, multi-currency and multi-tenant support lets banks run several entities or sponsor-bank programmes on one deployment.
Proven across markets. M2P's platform is used by 200+ banks and 300+ lenders across 30+ countries, including India, the Middle East, Africa and Southeast Asia.
Whether you are modernising a legacy core, launching a digital bank or preparing for AI-driven distribution, M2P can help you move from AI pilots to production with the controls regulators expect.
Talk to our core banking experts to see how Turing CBS can bring real-time data, embedded AI agents and built-in governance to your bank.
What is AI in core banking? AI in core banking means using machine learning, generative AI and AI agents on top of, and connected to, the core banking system. The core remains the system of record, while AI reads its data to detect fraud, assess credit, automate operations and serve customers, and sends instructions back through controlled interfaces.
Will AI replace the core banking system? No. AI depends on the core for accurate data and for executing transactions. A modern, real-time core becomes more important as AI adoption grows, because it enforces the limits, permissions and audit trails that keep AI decisions safe.
What are the most common AI use cases in banking today? Fraud and AML detection, credit scoring and early-warning systems, customer service assistants, document processing for KYC and lending, reconciliation and back-office automation, and developer productivity.
Who is liable if an AI system makes a wrong decision? The bank remains accountable to the customer and the regulator, whether it built the model or bought it. Banks can share financial exposure with vendors through contracts, but not the regulatory responsibility.
How does the RBI regulate AI in banking? The RBI's FREE-AI framework sets out seven principles and 26 recommendations for responsible AI. In June 2026, the RBI also released draft model risk management guidance proposing board-approved frameworks, human oversight and the ability to switch off AI models.
Is AI in credit scoring regulated in Europe? Yes. The EU AI Act classifies credit scoring as high-risk. Its high-risk obligations, including risk management, documentation and human oversight, apply from 2 December 2027.
What makes a core banking system AI-ready? Real-time, event-driven data; API-first design; composable product configuration; cloud-native scale; built-in controls and audit trails; and interfaces that let partners and AI agents discover and use banking products securely.